[Show abstract][Hide abstract] ABSTRACT: We observe that conventional approaches to the construction of likelihood models of visual appearance for image features are non-quantitative, precluding their use in tasks such as hypothesis testing for projected view validation. This document outlines a quantitative approach for verification of 3D objects' predicted edge features in images, which incorporates both the effects of image noise and local image structure. This approach supports the construction of a joint probability for the degree of conformity of image data to both edge orientation and location, without the need for arbitrary relative scale factors. The method has been validated on multiple views of man-made objects constructed froma variety of materials.
Computer and Robot Vision, 2008. CRV '08. Canadian Conference on; 06/2008
[Show abstract][Hide abstract] ABSTRACT: This paper describes an approach for the representation of projected 3D edge features for pur-poses of view-based recognition and localisation of objects. It is based upon the representation of arbitrary configurations of features using geometric co-occurrence as Pairwise Geometric Histograms (PGH). We describe a mathematical model for the interpolation of correlated changes in these histograms, comprising two independent linear models for use during simul-taneous view and scale matching. Mismatch and match distributions are provided to give a context for the accuracy of approximation. Assessment of the utility of this approach for object localisation is given in a companion paper at this conference.
[Show abstract][Hide abstract] ABSTRACT: This essay presents various factors which are suggested to warrant consideration for machine vision oriented PhD students. The factors suggest principles of best practice which are intended to ensure the validity of the work undertaken for the PhD. These ideas also have value in the assessment of research material when conducting literature surveys. The views reflect the methodologies upon which the TINA  open source computer vision system is founded.
[Show abstract][Hide abstract] ABSTRACT: A comparison of view-based and 3D model-based methods for localisation of man-made objects is made in the context of a working system. A projection validation approach is taken in order to confirm location hypotheses, which is based upon quantitative statistical models of feature detection and orientation. Results are provided which suggest that 3D data from stereo vision systems might be better employed for the prediction of novel views of objects, rather than as a generator of spatial representation suitable for geometric reasoning.